A robot-planning method that keeps multiple possible futures alive as other agents reveal their intentions recorded a 94% interaction success rate in a simulation study, compared with 66% for a conventional single-branch method. The authors describe that gap as a 42% relative increase. The result came from simulated three-agent road-crossing interactions.
The work is an arXiv version 1 preprint. Its central question was whether progressively branching contingency plans, paired with an ADMM solver, could outperform a conventional single-branch formulation in interaction performance and computation.
A plan that branches as uncertainty clears
Under the proposed architecture, sources of uncertainty about other agents' intentions are resolved at different branching times. Planned trajectories can then adapt as information arrives asynchronously.
The computational work is divided across possible scenarios with a method called the alternating direction method of multipliers, or ADMM. In plain terms, the solver handles scenario-dependent interaction games in parallel, then reconciles the local trajectories over the segments they share.
The main test used 100 paired receding-horizon simulations of a three-agent road-crossing scenario, with intent combinations represented equally. Every sampled instance was solved once under each branching structure with identical initial conditions, ground-truth intents, observation noise and random seed, so both methods faced the same sampled conditions.
The speed result was as striking as the success gap
The multi-branch planner also had a lower reported mean numerical solve time: 0.46 seconds per planning step, compared with 1.72 seconds for the single-branch version. That was an approximately 73% reduction, and the multi-branch method was faster in 97% of cases.
Among scenarios completed by both methods, the multi-branch formulation was reported to have 15% lower closed-loop cost and 26% lower variation in consecutive normalized ego controls, a measure of how much the robot's commands changed from step to step. Those figures are conditional on shared successful completion.
The study measured interaction success, final cost per executed step, RMS variation of consecutive normalized ego controls and mean solve time per simulation step. Empirical performance was reported with 95% confidence intervals for every metric.
Useful evidence, with sharp boundaries
The reported advantage was not universal. The paper says multi-branching did not strictly dominate every simulated interaction, and it identifies a failure mode near feasibility boundaries when ADMM had not converged sufficiently. Those cases were associated with safety violations in the simulation.
There is also a mathematical limit to what the solver result can establish. Because the trajectory games are nonconvex, standard convergence guarantees for convex ADMM do not apply. The local guarantee described by the authors does not ensure convergence from arbitrary starting points or show that the stationary point is a generalized Nash equilibrium.
The timing comparison has a practical qualification. It excludes symbolic construction, compilation and the initial cold-started solve, while solver structures were reused. The reported figures therefore cover the online solve stage under those conditions.
A larger demonstration points to the next test
A separate overtaking demonstration expanded the setup to four uncertain traffic agents and 16 joint intent scenarios. In that single proof-of-concept comparison, the multi-branch planner completed the maneuver safely after 58 simulation steps, with a median online solve time of 2.14 seconds, while the single-branch planner failed within the simulation horizon.
That was a single proof-of-concept demonstration rather than a repeated comparative evaluation. The study's evidence consists of the 100 paired three-agent simulations and this one overtaking setup, so its conclusions remain tied to the tested simulation cases.
Funding
The acknowledgements report partial support for several authors through the Architecture des Systèmes Complexes chair and for David Fridovich-Keil through the Hi! PARIS international visiting chair program.
Paper data and sources
Original title: Cascaded consensus splitting for multi-branch contingency games
Authors: Bastien Lechardoy, Pau de las Heras Molins, Thibault Lahire et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text